shithub: opus

--- a/dnn/nnet.c

+++ b/dnn/nnet.c

@@ -415,7 +415,7 @@

 #define MAX_CONV2D_INPUTS 2048

-void compute_conv2d(const Conv2DLayer *conv, float *out, float *mem, const float *in, int len2, int activation)

+void compute_conv2d(const Conv2dLayer *conv, float *out, float *mem, const float *in, int len2, int activation)

    int i;

    const float *bias;

--- a/dnn/nnet.h

+++ b/dnn/nnet.h

@@ -83,7 +83,7 @@

   int out_channels;

   int ktime;

   int kheight;

-} Conv2DLayer;

+} Conv2dLayer;

 typedef struct {

   const float *bias;

@@ -175,6 +175,7 @@

 extern const WeightArray rdovaeenc_arrays[];

 extern const WeightArray rdovaedec_arrays[];

 extern const WeightArray fwgan_arrays[];

+extern const WeightArray pitchdnn_arrays[];

 int linear_init(LinearLayer *layer, const WeightArray *arrays,

   const char *bias,

@@ -231,6 +232,8 @@

   int kernel_size,

   int nb_neurons,

   int activation);

+void compute_conv2d(const Conv2dLayer *conv, float *out, float *mem, const float *in, int len2, int activation);

 int embedding_init(EmbeddingLayer *layer, const WeightArray *arrays,

   const char *embedding_weights,

--- /dev/null

+++ b/dnn/pitchdnn.c

@@ -1,0 +1,61 @@

+#ifdef HAVE_CONFIG_H

+#include "config.h"

+#endif

+#include <math.h>

+#include "pitchdnn.h"

+#include "os_support.h"

+#include "nnet.h"

+#include "lpcnet_private.h"

+int compute_pitchdnn(

+    PitchDNNState *st,

+    const float *if_features,

+    const float *xcorr_features

+    )

+{

+  float if1_out[DENSE_IF_UPSAMPLER_1_OUT_SIZE];

+  float downsampler_in[NB_XCORR_FEATURES + DENSE_IF_UPSAMPLER_2_OUT_SIZE];

+  float downsampler_out[DENSE_DOWNSAMPLER_OUT_SIZE];

+  float conv1_tmp1[NB_XCORR_FEATURES + 2] = {0};

+  float conv1_tmp2[NB_XCORR_FEATURES + 2] = {0};

+  float output[DENSE_FINAL_UPSAMPLER_OUT_SIZE];

+  int i;

+  int pos=0;

+  float maxval=-1;

+  PitchDNN *model = &st->model;

+  /* IF */

+  compute_generic_dense(&model->dense_if_upsampler_1, if1_out, if_features, ACTIVATION_TANH);

+  compute_generic_dense(&model->dense_if_upsampler_2, &downsampler_in[NB_XCORR_FEATURES], if1_out, ACTIVATION_TANH);

+  /* xcorr*/

+  OPUS_COPY(&conv1_tmp1[1], xcorr_features, NB_XCORR_FEATURES);

+  compute_conv2d(&model->conv2d_1, &conv1_tmp2[1], st->xcorr_mem1, conv1_tmp1, NB_XCORR_FEATURES, ACTIVATION_TANH);

+  compute_conv2d(&model->conv2d_1, &conv1_tmp1[1], st->xcorr_mem2, conv1_tmp2, NB_XCORR_FEATURES, ACTIVATION_TANH);

+  compute_conv2d(&model->conv2d_1, downsampler_in, st->xcorr_mem3, conv1_tmp1, NB_XCORR_FEATURES, ACTIVATION_TANH);

+  compute_generic_dense(&model->dense_downsampler, downsampler_out, downsampler_in, ACTIVATION_TANH);

+  compute_generic_gru(&model->gru_1_input, &model->gru_1_recurrent, st->gru_state, downsampler_out);

+  compute_generic_dense(&model->dense_final_upsampler, output, st->gru_state, ACTIVATION_LINEAR);

+  for (i=0;i<DENSE_FINAL_UPSAMPLER_OUT_SIZE;i++) {

+    if (output[i] > maxval) {

+      pos = i;

+      maxval = output[i];

+    }

+  }

+  return (1.f/60.f)*pos - 1.5;

+  /*return 256.f/pow(2.f, (1.f/60.f)*i);*/

+}

+void pitchdnn_init(PitchDNNState *st)

+{

+  int ret;

+  OPUS_CLEAR(st, 1);

+  ret = init_pitchdnn(&st->model, pitchdnn_arrays);

+  celt_assert(ret == 0);

+  /* FIXME: perform arch detection. */

+}

--- /dev/null

+++ b/dnn/pitchdnn.h

@@ -1,0 +1,30 @@

+#ifndef PITCHDNN_H

+#define PITCHDNN_H

+typedef struct PitchDNN PitchDNN;

+#include "pitchdnn_data.h"

+#include "lpcnet_private.h"

+#define NB_XCORR_FEATURES (PITCH_MAX_PERIOD-PITCH_MIN_PERIOD)

+typedef struct {

+  PitchDNN model;

+  float gru_state[GRU_1_STATE_SIZE];

+  float xcorr_mem1[(NB_XCORR_FEATURES + 2)*2];

+  float xcorr_mem2[(NB_XCORR_FEATURES + 2)*2*8];

+  float xcorr_mem3[(NB_XCORR_FEATURES + 2)*2*8];

+} PitchDNNState;

+void pitchdnn_init(PitchDNNState *st);

+int compute_pitchdnn(

+    PitchDNNState *st,

+    const float *if_features,

+    const float *xcorr_features

+    );

+#endif

--- a/dnn/torch/neural-pitch/export_neuralpitch_weights.py

+++ b/dnn/torch/neural-pitch/export_neuralpitch_weights.py

@@ -52,7 +52,7 @@

     message = f"Auto generated from checkpoint {os.path.basename(args.checkpoint)}"

-    writer = CWriter(os.path.join(args.output_dir, "neural_pitch_data"), message=message, model_struct_name='PitchDNN')

+    writer = CWriter(os.path.join(args.output_dir, "pitchdnn_data"), message=message, model_struct_name='PitchDNN')

     writer.header.write(

 f"""

 #include "opus_types.h"

--

⑨